YNSScarSaiyan/bulmax-zip

BulmaX Zip — native thinker checkpoint

Release status (26 September 2026): This public repository is a weight archive for the verified recurrent-leap thinker readout. The first native COHR decoder run aborted in the Intel GPU driver before a quality report; the six-run Arc A750 campaign is pending a checkpointed restart. There is no released end-to-end question-answering model or final Test 2 quality verdict. The permanent 200-question held-out set has not been evaluated.

BulmaX Zip is an experimental language-system architecture by Dakuwon Moody / Saiyan Corp. Its current path combines a fixed sparse recurrent reservoir, an online readout, a local-update COHR graph, and an optional count-based MELF candidate router. The project aims at frontier-scale capability, but this checkpoint has not demonstrated that capability or a win over transformers.

What this archive contains

File	Content	SHA-256
thinker_rls_w_prod.safetensors	Native Test 2 readout, key w, float32, shape 20000 × 384, 30,720,080 bytes	9cf37d46c692aaa1ed87c686355f86fc3072108ea247f1f3b1dca8cdf0627773
checkpoint_manifest.json	Identity, scope, and checksums for this snapshot	See file

The reservoir is fixed and generated deterministically from the native recipe; its weight hash is 0b8a204667bc12260fa6e869775e96891c18713157dc1eade0f18ee8e722fecb. The readout tensor hash inside the verified production export is 644844371e5b4d156b78da2e07ae796f294b47f46ed3fa65b8ce92ef0c8b1c29. The archive does not contain the third-party MiniLM weights, the QA training/validation streams, or an unfinished COHR decoder checkpoint.

The source mirror is private. It is mirrored from the native-only project repository and includes the full experiment contracts and verification receipts. The released readout is a research checkpoint, not a standalone service.

Current architecture and training

• Host encoder: frozen MiniLM-L6-v2 ONNX graph, masked mean pooling and normalization. The input question also drives the reservoir through its first 48 raw word-embedding rows, including special tokens and excluding padding.
• Recurrent state: 20,000 units, approximately 1% sparse recurrent connectivity, spectral radius 0.9, leak 0.3, input scale 1.0. Native NumPy PCG64 seed 0 generates the fixed weights; oneAPI SYCL/ESIMD executes the Arc A750 reservoir path.
• Readout: 20,000 × 384 sliding-window block RLS, block size 128, window 2,000, delta 0.05. It learned from 4,000 weighted QA draws to predict a normalized answer-minus-question embedding direction. The reservoir and encoder remained frozen; this path used no backpropagation.
• Decoder under evaluation: COHR candidate motor with local plasticity, compared with COHR+MELF. A pinned GPT-2 BPE tokenizer provides token IDs only; no GPT-2 model parameters or pretrained logits enter the native decoder. Production decoder math runs on the Intel Arc A750.

The implementation uses ONNX Runtime on the host, NumPy/SciPy for reference checks, and oneAPI SYCL/ESIMD for the accelerator. Its current code and checkpoint do not require PyTorch.

Verified evidence and its limits

The native Test 1 v3 long-memory prerequisite passed its frozen three-seed recall gate at distances 2, 4, 8, and 16. Its smallest recurrent-minus-static gap was +0.069705, above the +0.05 threshold. An earlier v2 gradient-free probe failed at distances 8 and 16; v3 tested a separately frozen analytic readout and does not erase that result.

The 4,000-step Test 2 thinker export passed an identity and integrity audit: 20,000 training draws from 2,306 distinct questions, 2,000 distinct validation questions, no question overlap, and a 20000 × 384 saved readout. Training MSE was 0.0030362944. A 500-question diagnostic found mean cosine 0.08211634 between the prediction and answer embedding. On all 2,000 validation questions, the prediction's mean cosine to the answer-minus-question direction was 0.25373 for the matched question versus 0.09612 when rotated to another question (paired gap +0.15761). This shows a question-specific leap signal; it does not measure decoded answer quality. Direct answer-embedding cosine was slightly worse for matched than shuffled predictions (0.07332 versus 0.08316).

Six Arc parity checks passed for the native candidate, motif, plasticity, and hidden-state paths. The production decoder campaign is designed to evaluate three seeds for COHR-only and COHR+MELF on the frozen 2,000-question validation split. The predeclared gate requires true thinker vectors to improve full-vocabulary teacher-forced NLL by at least 0.05 nats/token against both shuffled and zeroed controls in every seed. A separate 0.05 gate tests whether the hybrid improves over COHR-only. Exact match, token accuracy, coverage, truncation, state size, and runtime are also reported. No result from this campaign is asserted until complete reports pass audit.

The earlier framework-based Phase 2b decoder experiments failed their preregistered gate. Those historical results are not evidence for the current native decoder and their weights are not part of this archive.

Intended use and limitations

This archive preserves a verified experimental readout while the full system is being evaluated. It is suitable for provenance, comparison, and continued BulmaX Zip research. It is not a finished QA model and should not be represented as one. The current training pool has only 2,306 distinct questions, the decoder gate is pending, and no independent broad benchmark or matched transformer baseline has been run. The permanent held-out set remains sealed until the frozen model and gate decisions are fixed.

The frozen MiniLM encoder is itself a pretrained transformer dependency. A future claim about outperforming transformers must account for that external representation source and compare end-to-end systems under a fair protocol.

Provenance

• Frozen protocol: runs/t2_run_card_v3.md in the private source mirror.
• Production thinker receipt: runs/t2_native_production_export_20260926.json.
• Native long-memory gate: runs/t1_native_v3_gate_final_20260926.json.
• Leap-signal analysis: runs/t2_production_thinker_geometry_20260926.json.
• Public weight file hash and internal readout hash are recorded above and in checkpoint_manifest.json.

No license has been declared for this research checkpoint.